Uninformative Input Features and Counterfactual Invariance: Two Perspectives on Spurious Correlations in Natural Language

Anonymous

Abstract: The natural language processing community has become increasingly interested in spurious correlations, and in methods for identifying and eliminating them. Gardner et al (2021) argue that due to the compositional nature of language, \emph{all} correlations between labels and individual input features are spurious. This paper analyzes this proposal in the context of a toy example, demonstrating three distinct conditions that can give rise to feature-label correlations through a simple PCFG. Linking the toy example to a structured causal model shows that (1) feature-label correlations can arise even when the label is invariant to interventions on the feature, and (2) feature-label correlations may be absent even when the label \emph{is} sensitive to interventions on the feature. Because input features will be individually correlated with labels except in very rare circumstances, mitigation and stress tests should focus on those correlations that are counterfactually invariant under plausible causal models.

Paper Type: short